BIRD-SQL Qwen3.5-4B Generator
This repository contains the deployable checkpoint used as the SQL candidate generator in the BIRD-SQL submission pipeline at https://github.com/AeBoPi/bird-test-submission.
Model
- Architecture: Qwen3.5-4B
- Mode: non-thinking
- Checkpoint:
checkpoint-372 - Training data: 16,830 SFT records obtained by duplicating 8,415 processed BIRD-train shell-agent traces
- Output role: generate SQL candidates; the separately trained selector is available at https://huggingface.co/symbolzh/selector0827
The included chat_template.jinja, tokenizer files, and generation
configuration are part of the frozen deployment artifact and should be kept
with the weights.
Usage
The complete data preparation, vLLM startup, candidate generation, selector,
resume, and output commands are documented in the submission pipeline
repository. A typical local vLLM deployment points --model directly to the
downloaded directory.
This model is intended for research on BIRD-SQL. Generated SQL can be invalid or unsafe for arbitrary databases. The accompanying pipeline restricts each task to its own SQLite database and applies a read-only SQL policy.